The New Math of AI Venture Capital

Venture capital built its modern playbook on a single structural insight: software has near-zero marginal costs. Once you write the code, each additional user costs almost nothing to serve. That simple fact, compounded over decades, produced gross margins above 80 percent, the CAC/LTV ratio as a north star, and the recurring-revenue model as the closest thing to a guaranteed return that private markets have ever seen. It also produced a generation of funds that trained their partners to spot one pattern and deploy enormous capital against it. That playbook is now under serious pressure, and the source of the pressure is the same technology those funds have spent the last three years celebrating.

The Three-Pillar Model

The software VC thesis rested on three assumptions that held, with remarkable consistency, from roughly 1995 through 2022. First: marginal cost approaches zero as the business scales, so every new dollar of revenue is nearly pure margin. Second: the CAC/LTV ratio, properly constructed, predicts whether a company can grow profitably — if you spend less acquiring a customer than you extract in lifetime revenue, the math works. Third: recurring subscription revenue creates predictability, which justifies high entry valuations because the compounding is visible and can be modeled years in advance.

These three pillars reinforced each other. Low marginal cost made LTV high. High LTV justified aggressive CAC spending. Recurring revenue made both predictable enough to price. The result was a category where the best companies could grow at 100 percent annually while improving margins, and growth itself became self-funding once the acquisition engine was tuned. The software-as-a-service model was so reliably profitable at scale that investors learned to look past early losses and price the terminal state — a discipline that worked staggeringly well for two decades. AI companies have demolished all three pillars simultaneously, and the investors who haven't rebuilt their frameworks are evaluating fundamentally different businesses using a lens designed for a different era.

How AI Breaks the First Pillar

The compute cost problem is the most immediate issue, but its full implications are still underpriced in most venture models. Every inference call — every time a user asks an AI product to do something — costs money. Tokens are not free. Inference is not free. At scale, compute becomes a meaningful fraction of revenue, and unlike traditional software infrastructure, it does not decline as a percentage of revenue simply because the company grows. It falls in absolute terms as models become more efficient, but heavy usage creates heavy costs, and heavy users are often a company's best customers.

The gross margin consequence is stark. Enterprise SaaS businesses built on traditional software routinely deliver 75-85 percent gross margins at scale. AI-native companies building on foundation models are running 50-70 percent gross margins in optimistic scenarios, and considerably lower during rapid scaling when inference costs haven't been engineered down. This is not a temporary growing pain — it reflects a structural difference in the cost of delivering an AI-powered outcome versus delivering a data query or a rendered page. The companies that will eventually look like their software predecessors are the ones that train their own specialized models at lower inference cost or move computation to the edge to escape per-token cloud pricing. Most haven't done either yet, and current valuations rarely reflect that journey's cost.

The LTV Formula Breaks Under Usage Pricing

The second pillar collapses in a subtler way. The SaaS lifetime value formula assumes that the cost to serve a customer is roughly fixed — once the product is built, all users cost about the same to support. Usage-based pricing, which nearly every AI company has adopted because it mirrors how compute costs actually work, destroys this assumption entirely. When a customer uses an AI product more heavily, they generate more revenue but also more cost, and the contribution margin of a heavy user is not obviously better than that of a light user, and may be worse if the use case is inference-intensive.

This creates a troubling inversion of standard SaaS logic. The customers who seem most engaged — driving the highest usage numbers — may be generating the lowest contribution margins. Traditional SaaS engagement metrics (daily active users, session depth, feature adoption) are proxies for LTV because more engagement predicts less churn and higher expansion revenue. In an AI context, engagement also predicts more compute spend, and the relationship between engagement and profitability is no longer monotonically positive. Investors who haven't rebuilt their LTV models for this cost structure are making decisions on inputs that don't measure what they think they measure, and the capital allocation errors that follow are compounding.

The Valuation Divergence That Doesn't Close Quietly

Private market valuations for AI companies have reached multiples that require the companies to become some of the most dominant software businesses ever built just to close the math. What is underappreciated is the mechanism by which these valuations can persist without ever resolving: the same institutional pressure that inflated them also creates reluctance to mark them down, subsequent rounds paper over the problem with fresh capital, and secondary market liquidity creates enough exit opportunities for early investors to realize returns before the reckoning arrives. The companies most exposed are those whose advantage is primarily model performance at a moment when foundation model capability is converging across providers, a dynamic covered in depth in AI valuations.

The problem is not that every AI company is overvalued — some will justify extraordinary prices. The problem is that the distribution of outcomes in AI is fat-tailed in a way that makes current pricing incoherent at the portfolio level. A basket of AI startups valued at 50x forward revenue is a bet that at least one will become a dominant platform, which is reasonable at the portfolio level but a poor bet in any individual investment. The same investors who understand this at the portfolio level are often making individual deployment decisions that contradict it.

Competitive Landscape

The competitive dynamics in AI investing are unlike any prior technology cycle because the largest potential competitors are also the infrastructure providers. Foundation model companies — OpenAI, Anthropic, Google DeepMind, Meta — simultaneously power the startup ecosystem and compete with it by extending their own products toward every high-value use case. A startup that builds a legal AI tool on top of a foundation model API is one product announcement away from competing with the model provider itself.

Hyperscalers occupy an even more structurally advantaged position. AWS, Azure, and Google Cloud profit from every inference call made by every AI startup on their infrastructure, regardless of which startup wins. They have no incentive to pick a winner — they are the toll road that every winner must pay. This means the economic moats most worth paying for in AI are those that escape the foundation model and cloud infrastructure trap: proprietary data pipelines that can run on any model, workflow integrations with switching costs that predate the current model generation, and distribution scale that generates compounding data advantages over time.

Investment Thesis

The vertical AI thesis has become the most credible framework for finding durable returns in this environment. Vertical integration in regulated industries — healthcare, legal, financial services — creates compliance-based lock-in that outlasts model improvements, because the value being delivered is not raw model capability but the surrounding workflow, the regulatory fit, and the trained compliance staff that depend on the system. Companies with proprietary clinical records, legal precedent databases, or financial history that cannot be replicated through web scraping have a data layer underneath their AI layer that is genuinely scarce.

Distribution scale at the consumer level creates a second category of defensibility. A product with tens of millions of daily active users generating interaction data can retrain more efficiently and more relevantly than any competitor, including the foundation model providers themselves. The interaction data becomes a proprietary fine-tuning corpus, and the user relationship becomes the asset. The common thread in both categories is that the defensible companies are not selling model performance — they are selling outcomes in specific contexts where they have structural advantages that survive the inevitable commoditization of foundation model capability itself.

Limitations

This analysis rests on the assumption that foundation model performance will continue to converge across providers, reducing the advantage of any company whose moat is primarily access to a more capable model. If a single provider achieves a durable capability lead — whether through a breakthrough in reasoning, multimodality, or efficiency — the competitive dynamics change in ways this framework doesn't fully capture. Similarly, regulatory intervention could reverse the current architecture: if inference costs are mandated to stay above a floor, or if data-sharing requirements undermine proprietary dataset moats, the relative position of different AI business models shifts substantially. The framework is built on current trends, and current trends are moving fast.

The Bottom Line

The venture capital industry is running the most consequential experiment in its history: deploying more capital into a single technology category, faster, at higher prices, than it ever has before. Some of that capital will produce extraordinary returns. Most of it will not — not because the technology is failing, but because the economic model of AI companies is different enough from the SaaS companies that trained investors' intuitions that standard comparisons mislead more than they inform. The investing lesson is the same one every technology cycle eventually delivers: understand the actual economics before assuming the last playbook still applies.

The funds that build new analytical frameworks — that price compute costs into LTV, that distinguish data moats from model dependencies, that understand the difference between distribution advantages and temporary performance leads — will compound well from this cycle. The ones that apply 2015 SaaS metrics to 2026 AI businesses are making a category error that the market will eventually correct, loudly. The next wave of returns will come from companies that solved the compute cost problem or built moats the models can't replicate — not from model performance alone.

References

• Bessemer Venture Partners, State of the Cloud, bessemervp.com — annual benchmarking of SaaS and cloud business metrics.
• Sequoia Capital, Generative AI's Act Two, sequoiacap.com — analysis of the AI application layer and where value accrues.
• a16z, The New Economics of AI, a16z.com — framework for understanding gross margin compression in AI-native businesses.

Related

• AI Valuations Are Detaching — how private market pricing is diverging from cash-flow fundamentals across the AI sector.
• The Vertical AI Thesis — why domain-specific models with proprietary data are the durable bet as foundation models commoditize.
• The Data Moat in the AI Era — how proprietary data is becoming the last defensible position when model capability converges.